Automated soybean quality detection system using deep learning

Accurate quantification of impurity and crushing rates during soybean harvesting is critical for intelligent combine harvester optimisation. This research presents an integrated detection system that synergises mechanical monolayer sampling with an enhanced scSE-UNet model to address this challenge. A purpose-built imaging device was developed to capture standardised soybean kernel images through controlled monolayer deposition and uniform lighting conditions. The UNet architecture was enhanced by integrating a VGG16-based encoder for multi-scale feature extraction and scSE-block+ modules at decoder skip connections to amplify discriminative spatial-channel features, combined with a Dice-Focal hybrid loss function to mitigate class imbalance caused by kernel adhesion and complex backgrounds. Experimental validation achieved segmentation accuracies of 97.8 % (whole kernels), 97.2 % (broken kernels), and 96.3 % (impurities), with a mean IoU of 93.9 %. The system demonstrates operational efficiency with 0.153-second/image processing latency, delivering final detection accuracies of 89.8 % for impurity rate and 88.2 % for crushing rate, confirming its robustness for real-time harvester monitoring applications.